config.yml

data:
  synthetic:
    data_type: nerf_synthetic
    dataset: lego # options : chair, drums, ficus, hotdog, lego, materials, mic, ship
    run_type: train # options : train, val ,test

    factor: 8
    no_ndc: True # 'do not use normalized device coordinates (set for non-forward facing scenes)
    spherify : True

    path_zflat: False # regarding render_poses(poses for test img), choose spiral or path
    bd_factor: .75
    recenter : True
    llffhold: 8 # will take every 1/N images as LLFF test set, paper uses 8

    half_res : True
    white_bkgd: False # set to render synthetic data on a white bkgd (always use for dvoxels)

  llff:
    data_type: nerf_llff_data
    dataset: fern # options : fern, flower, fortress, horns, leaves, orchids, room, trex
    run_type: train # options : train, val ,test

    factor: 8
    no_ndc: False # 'do not use normalized device coordinates (set for non-forward facing scenes)
    lindisp: False # sampling linearly in disparity rather than depth
    spherify : False # Should be synchronized with Image source camera position distribution

    path_zflat: False # regarding render_poses(poses for test img), choose spiral or path
    bd_factor: .75
    recenter : True
    llffhold: 8 # will take every 1/N images as LLFF test set, paper uses 8


  thin_structure:
    data_type: neus_thin_structure
    dataset: thin_catbus

    camera_outside_sphere: True
    scale_mat_scale : 1.1

    white_bkgd: False

embed:
  nerf:
    include_input: True
    input_dims : 3
    log_sampling: True

  neus:
    include_input: True
    input_dims : 3
    log_sampling: True

  hypernerf:
    # NeRF metadata configs.
    use_nerf_embed: False
    nerf_embed_key: appearance
    nerf_num_embeddings: 8 # TODO
    nerf_num_dims: 8 # TODO

    # Warp configs.
    use_warp: False
    warp_embed_key: warp
    warp_num_embeddings: 8 # TODO
    warp_num_dims: 8 # TODO

    warp_field_cls: Callable[..., nn.Module] = warping.SE3Field

    use_alpha_condition: bool = False
    use_rgb_condition: bool = False

    hyper_slice_method: none
    hyper_num_embeddings: 8 # TODO
    hyper_num_dims: 8 # TODO
    hyper_embed_cls: Callable[..., nn.Module] = (
      functools.partial(modules.GLOEmbed, num_dims=8))
    hyper_embed_key: str = 'appearance'

    hyper_use_warp_embed: bool = True
    hyper_sheet_mlp_cls: Callable[..., nn.Module] = modules.HyperSheetMLP
    hyper_sheet_use_input_points: bool = True

rendering:
  nerf:
    N_samples: 64 # number of coarse samples per ray
    N_importance: 64 # number of additional fine samples per ray
    perturb: 1. # set to 0. for no jitter, 1. for jitter
    use_viewdirs: True # use full 5D input instead of 3D
    i_embed: 0 # set 0 for default positional encoding, -1 for none

    multires: 10 # 'log2 of max freq for positional encoding (3D location)'
    multires_views: 4 # 'log2 of max freq for positional encoding (2D direction)'
    raw_noise_std: 1 # 'std dev of noise added to regularize sigma_a output, 1e0 recommended'

    render_only: False # do not optimize, reload weights and render out render_poses path
    render_test: False # render the test set instead of render_poses path
    render_factor: 0 #downsampling factor to speed up rendering, set 4 or 8 for fast preview

    # Camera Movements
    render_pose_num : 1 # default 120, newly added for factorization
    N_rots : 2 # Number of rotation for spiral path
    zrate: .5 # Moving speed along with z axis for spiral path

    # For deeepvoxel
    shape: greek # options, armchair/ cube/ greek/ vase

  neus:
    n_samples: 64 # The number of samples in z axis
    n_importance: 64 # Determine whether to up sampling or not
    n_outside: 0 # default with mask. without mask, use 32
    up_sample_steps: 4     # 1 for simple coarse-to-fine sampling. The number of updating z_vals
    perturb:  1.0 # positive value -> perturbation on samples of z values.

    render_pose_num: 1

  hypernerf:
    # NeRF rendering.
    num_coarse_samples: 196
    num_fine_samples: 196
    use_stratified_sampling: True
    use_white_background: False
    use_linear_disparity: False
    use_sample_at_infinity: True

    spatial_point_min_deg: 0
    spatial_point_max_deg: 10
    hyper_point_min_deg: 0
    hyper_point_max_deg: 4
    viewdir_min_deg: 0
    viewdir_max_deg: 4
    use_posenc_identity: True

    alpha_channels: 1
    rgb_channels: 3
    activation: nn.relu
    norm_type: None
    sigma_activation: nn.softplus


model:
  nerf:
    layer_num: 8 # Original netdepth: 8 # layers in network
    hidden_dim : 256 # netwidth: 256 # channels per layer
    layer_num_fine : 8 # netdepth_fine: 8 # layers in fine network
    hidden_feature_fine : 256 # netwidth_fine: 256 # channels per layer in fine network
    batch_size: 1024 # 32*32*4=4096, batch size (number of random rays per gradient step)

    lrate: 5.0E-4 # learning rate
    lrate_decay: 250 # exponential learning rate decay (in 1000s)
    lrate_schedule_gamma: 0.1
    N_iters: 200000 # 2500 * 4096(N_rand) = 5GB GRAM usage

    chunk: 32768  # 65536, 1024*32, number of rays processed in parallel, decrease if running out of memory. (Rendering Image) 4585 - 1985 = 2600MiB Uses
    netchunk: 65536 # 1024*64, number of pts sent through network in parallel, decrease if running out of memory
    no_batching: False # only take random rays from 1 image at a time -> latency at ray_generation ~10s
    no_reload: False # do not reload weights from saved ckpt
    ft_path: None # specific weights npy file to reload for coarse network
    random_seed: None, # fix random seed for repeatability

    precrop_iters: 0 # 'number of steps to train on central crops'
    precrop_frac: .5 # fraction of img taken for central crops

  neus: # with mask configuration
    learning_rate: 5.0E-4
    learning_rate_alpha: 0.05
    N_iters: 300000

    batch_size: 16 # default 512
    validate_resolution_level: 4
    warm_up_end: 5000
    anneal_end: 50000 # default for with mask. without mask, wmaks : 0 , thin structure : 50000

    igr_weight: 0.1
    mask_weight: 0.0 # wmask : 0.1, thin structure : 0.0

    layer_num_fine: 8 # netdepth_fine: 8 # layers in fine network
    hidden_feature_fine: 256 # netwidth_fine: 256 # channels per layer in fine network

    layer_num: 8
    hidden_dim: 256

    backbone:
      D: 8
      d_in: 4
      d_in_view: 3
      W: 256
      multires: 10
      multires_view: 4
      output_ch: 4
      skips:
        - 4
      use_viewdirs: True

    sdf_network:
      d_out: 257
      d_in: 3
      d_hidden: 256
      n_layers:  8
      skip_in:
        - 4
      multires: 6
      bias: 0.5
      scale: 3.0 # used thin structure : 3.0 wmask : 1.0
      geometric_init: True
      weight_norm: True

    variance_network:
      init_val: 0.3

    rendering_network:
      d_feature : 256
      mode: idr
      d_in: 9
      d_out: 3
      d_hidden: 256
      n_layers:  4
      weight_norm: True
      multires_view: 4
      squeeze_out: True

  hypernerf:
    use_viewdirs: True
    noise_std: None

    # NeRF architecture.
    in_feature : 10 # TODO: dummy
    use_fine_model : True
    nerf_trunk_depth: 8
    nerf_trunk_width: 256
    nerf_rgb_branch_depth: 1
    nerf_rgb_branch_width: 128
    nerf_rgb_channels: 3
    nerf_alpha_depth: 0
    nerf_alapha_width: 8
    nerf_alpha_channels: 1
    nerf_norm : None
    nerf_activation: relu
    nerf_skips:
      - 4

    hypersheet:
      in_channels: 10 # TODO
      out_channels: 3 # TODO
      min_deg: 0
      max_deg: 1
      depth: 6
      width: 64
      skips:
        - 4
      use_residual: False
log:
  i_print: 200
  i_img: 5000 # frequency of tensorboard image logging, default 5000
  i_weights: 10000 # frequency of weight ckpt saving, 10000
  i_testset: 5000 # frequency of testset saving, 5000
  i_video: 20000 # frequency of render_poses video saving, 20000